Author: Merkle3s Capital; Source: X, @Merkle3sCapital

We have answered this question three times

Is there a bubble in AI?

This is the most frequently asked question in the past two years, and we have written about it more than once. Every time a conclusion is given, every time it is beaten back by new surges and plummets, it is re-examined.

We're not going to give you a simple "yes" or "no" answer this time.

Because this question itself is the wrong question. AI is not an asset, it is an entire industry chain - from wafer factories to power plants, from giants with a market capitalization of trillions to startups that have just received financing. Asking "Is there a bubble in AI?" is as crude as asking "Is there a bubble in real estate?": Can the core areas of first-tier cities and ghost towns in tier-18 counties have the same answer?

If you apply one question to all levels, the answer you get is bound to be wrong.

The correct question is:Which level is the AI bubble at?

We never ask "whether there is a bubble", only "where it is and how thick it is".

If you break this problem apart, you will see a picture that is counterintuitive: the layer where everyone is focused on and worried is the safest; and where the real bubble is, few people seriously discuss it.

Ghosts of 2000: What’s different this time

When talking about the AI bubble, we cannot avoid the year 2000. Butmost people only remember "the Internet bubble burst" and don't remember how it burst.

The script of that year: first find the stock price, then find the income

The script for the collapse in 2000 went like this: Telecom companies borrowed huge amounts of debt and frantically laid out fiber optics, like building an eight-lane highway in an empty city. The road is finished, where is the car? No. 85% to 95% of the optical fiber that was laid back then was "dark" - lying underground without a single bit being transmitted. The assets are on the books, the income is zero, and the debt is real. Then, boom.

Fiber optics is just the story of the infrastructure layer. The application layer is even more ridiculous.

The most famous pet supplies e-commerce company at that time had only a few million US dollars in annual revenue the year it went public, and its marketing expenses were several times its revenue. It spent money on advertising in the Super Bowl and lost one on every order it sold. The more it sold, the faster it lost money. About nine months after listing, it went into liquidation and collapsed. This is not an exception. This is the standard portrait of the application layer back then: zero profit, relying on financing to survive, and using "eyeballs" and "clicks" instead of revenue to value oneself.

What’s even more magical is that some scholars at that time made statistics: as long as a company changes its name and adds ".com" to the end, its stock price will rise a lot on average without changing any business.

The market is paying for suffixes, not business.

Look again at the "shovel seller" back then. Cisco is the NVIDIA of 2000 - all Internet traffic goes through its routers, and the logic is impeccable. But at the peak of the bubble, Cisco's price-to-earnings ratio hit triple digits. What concept? This means that the market requires it to maintain the profit scale at that time for more than a hundred years, or to increase it more than ten times in a few years, before the transaction can be considered as a return on investment. Later, the Internet really changed the world, and traffic really exploded—it took more than two decades for Cisco's stock price to return to its 2000 high.

Remember this case, it is the most important footnote in the whole article: The biggest tragedy of the year was not buying a fake company, but buying a real company at a hundred times the price.

The current script: income first, then increase stock price

Now cut to 2026.

No GPU is dark. Every chip produced is inserted into the rack the moment it comes off the assembly line, and is fully loaded with tokens in exchange for real money. It's not that the utilization rate is high, it's 100%. It's that customers can't buy it even if they queue up with money.

What about the application layer? Compare it to a large head model company. The annualized revenue of a leading player, which was less than $100 million 18 months ago, is now $45-47 billion, and it has achieved quarterly profits. Management originally planned to grow by 10 times, but actually grew by 80 times.

Put the leading companies of the two eras together:

• That year: several million in revenue, tens of millions in losses, and went bankrupt nine months after listing

• Now: Income has doubled hundreds of times in 18 months, and you have started to make money

Companies back then relied on "stories" to ask for money from the capital market; today's leading companies rely on contracts to collect money from customers. This is not a difference in degree, it is a difference in business model.

"Shovel Seller" has also changed its valuation logic. Today, Nvidia's price-to-earnings ratio is around thirty times -- only a fraction of Cisco's peak. What supports this valuation is not the imagination of the future, but the backlog of orders that have been signed and written into the production schedule.

In those days, we had to have the stock price first, then look for the income, and we would find death; now we have to have the income first, and then increase the stock price, so that we can catch up. The order is different, the ending is different.

The buyer has also changed. In 2000, it was telecommunications companies that borrowed debt to lay fiber optics; today, those who buy computing power are Microsoft, Google, Meta, and Amazon—the four companies with the richest cash flows on the planet, and they spend their own earned money.

In 2000, borrowed money was used to buy assets that no one was using; in 2026, earned money was used to buy assets that were not enough - these are two species!

But there is a crack in the wall

At this point, we must put on the brakes.

This "own cash flow" story is starting to fray at the edges. The total capital expenditures of the four major cloud vendors this year are approximately US$725 billion, a year-on-year increase of 77%. What size is this? Roughly equivalent to an entire year's GDP of a moderately developed country, smashed into data centers.

What’s even more eye-catching is Amazon: free cash flow plummeted from $26 billion to $1.2 billion, almost to zero, and long-term debt is climbing. In other words, the money the giants earned themselves was almost no longer enough to burn, so they began to borrow.

This is not a sign that the bubble has burst - the giants' balance sheets are still the strongest in the history of human business. But it is the first crack in the wall:"Cash flow buyer", the hardest logic of this round, is slipping from "completely established" to "generally established".

Worth a glance every quarter.

Finish the review of 2000. The biggest misleading left by that bubble to future generations is that everyone remembers that "the story is false", but forgets that what really kills the market is the out-of-control supply: No matter how true the story is, as long as everyone on the supply side can infinitely increase leverage to expand production capacity, excess will be a matter of time, and collapse will be a mathematical problem. On the other hand, the key to judging whether this round will repeat the same mistakes is not how attractive the story on the demand side is, but whether anyone on the supply side can apply the brakes.

This leads to the next question: Who is applying the brakes this round?

Publish the map first, then clear mines layer by layer: the five-layer pyramid of AI computing power

Before naming names one by one, first draw the entire industry chain. The AI computing power industry chain can be divided into five layers from bottom to top:

Say it again using a table:

This picture has a pattern that can be seen at a glance: The closer it is to physics, the less bubbles; the closer it is to the story, the more bubbles it has.

At the L0 level, it takes three to five years to expand production and tens of billions of dollars to build a factory. Even if you want to blow up the bubble, you can't blow it up - the supply is not cooperative at all. The further you go up, the looser the physical constraints and the greater the narrative space: when you reach the long tail of L4, a PPT can raise funds, and bubbles naturally gather there.

The only exception is the L2 interconnection layer - it is obviously hardware and should be protected by physical constraints, but it has become the place with the strongest smell of bubbles. Why? It will be dismantled later.

The first step in judging the AI bubble is not to look at market sentiment, but to see clearly which level of the pyramid you are on.

In this map, the reason why the L0 layer dares to directly mark "No Bubble" is because it is locked by two physical locks. Let’s talk about locks first, and then clear mines layer by layer.

The first lock: TSMC

Why do we judge that this round of AI capital expenditure will not get out of control? The answer is not on the demand side, but on the supply side.

There is one necessary condition for a bubble to burst:Excess supply. Tulips must be planted everywhere, fiber optics must be laid so that no one uses them, and houses must be built so that they cannot be sold. No glut, no crash. The real culprit of the disaster in 2000 was not that the Internet story was wrong, but that the supply of optical fiber was completely out of control—any telecommunications company could borrow money to dig trenches and bury cables, and no one could step on the brakes.

The supply of AI computing power is in the hands of a group of the most conservative people in the world.

The "Central Bank" in the AI Era

TSMC’s market share in advanced processes exceeds 90%, with a lead of about 9 to 15 months over Intel and Samsung, and the gap shows no sign of closing at the most advanced 2nm. This means one thing: the production of global AI chips is not determined by the market, but by TSMC.

It is like a central bank in the AI era - the Federal Reserve controls how much money is printed, and TSMC controls how much computing power is printed. The Federal Reserve needs to hold meetings, vote, and face political pressure to raise interest rates; TSMC controls the supply of computing power and only needs to not nod on its production expansion plan.

The presidents of this "central bank" are a group of senior engineers in their seventies who have experienced the events of 2001 and 2008. They see themselves as the guardians of the founder's legacy, and they have seen with their own eyes how the semiconductor bubble blew up and buried the entire industry. In their memory, the "slump after the surge" is not a textbook case, but an employee who has been laid off with their own hands and a production line that has been shut down with their own eyes.

So when Jen-Hsun Huang came to the door and asked to double or even triple production capacity - they refused.

Think about how counterintuitive this is: The hottest company on earth comes to you with unlimited orders and cash, begging you to expand production, and you say no. Only one company in the world can say this "no", and only one company has the final say.

By the way, a detail: Jen-Hsun Huang has cooperated with TSMC for more than 30 years and has never signed a formal procurement contract. It's all about shaking hands. This is not a management loophole, this is a system accumulated by thirty years of trust - and this is why TSMC dares to say "no" to its largest customer, but the largest customer can only accept it.

How tight is this lock

Digital level:

• The most advanced 2nm process, all production capacity will be sold out by the end of this year, and there will be no one left

• Kaohsiung is building five 2-nanometer wafer fabs at the same time - the largest parallel construction of advanced process fabs in human history. However, it takes three to five years from the start of construction to mass production of an advanced wafer fab, with an initial investment of more than 20 billion US dollars

• Even if we continue to build like this, by 2030, the monthly demand for 2 nanometers is expected to be 400,000-450,000 pieces, and the production capacity will only be 300,000-350,000 pieces - a long-term gap of 100,000-150,000 pieces/month, which is equivalent to a quarter to a third of the demand that will never be met

There is a more hidden bottleneck: advanced packaging. Chips are only semi-finished products, and computing chips and memory must be "packaged" together before they can be used. This is the "last mile" of AI chips, and this road is basically guarded by TSMC, and production capacity is also in short supply all year round.

If TSMC completely frees up its hands, Nvidia could theoretically ship $2 to $3 trillion in GPUs a year—a figure that is nearly ten times the current actual shipments. It is TSMC that has locked this number.

All the world’s AI ambitions combined will line up in front of TSMC’s production capacity table.

This lock can also be picked

For the sake of fairness, let’s also state the opposite side clearly. This lock is not a perpetual motion machine, it has a script for being pried open: if someone - whether it is a Musk-like madman or Intel eager to turn over - bypasses TSMC, builds its own super fab cluster with the support of equipment manufacturers, and breaks the monopoly of advanced production capacity, then the discipline of production expansion will collapse.

By then, every chip factory will be like telecommunications companies in 2000, frantically building capacity, and the engine of oversupply will really ignite.

The good news is: the physical cycle of building the factory is there, and there is almost no possibility of this script being staged before 2027. The bad news is: once filming begins on this script, there will be no trailer.

Bubbles require runaway supply. And the AI ​​supply valve is in the hands of a group of old people who have seen two crashes and rejected Huang Renxun!

The second lock: electricity

Even if TSMC figures out the crazy expansion of production tomorrow, there still has to be a place to insert the chips after they are manufactured.

This is the second lock: electricity and land.

Many people think that the bottleneck of AI infrastructure is chips. In fact, the real bottleneck at the moment is something more earthly—Land approval and power grid access for data centers.

The absurdity of this matter is the mismatch of time scales. It takes two years to design a chip; it takes two or three years to build a data center; but to provide enough power for a data center - building a new power plant, expanding substations, pulling high-voltage transmission lines, completing environmental impact assessments and approvals - can easily take five years. Chips discuss nanometer evolution, and power grids discuss the ten-year plan.

Chips are updated in months, and power grids are updated in ten years—this is the biggest time difference in the AI era.

So you will see a strange scene: technology giants with tens of billions of dollars in budgets are looking for "land with electricity" all over the world, just like gold diggers looking for water. Buy land next to a nuclear power plant, sign a 20-year power purchase agreement, and even directly pay to restart decommissioned nuclear reactors. Money is not the problem, electricity is.

The power gap is not expected to be gradually alleviated until 2027-2028 - the construction cycle of power plants and power grids determines this timetable, and no amount of money can compress it much.

The effect of two locks stacked together is: the growth of AI computing power is forcibly "flattened". Demand wants to explode, and supply can only climb. As a result, growth has become slower, but also longer and more stable - this is precisely the treatment that technological revolutions such as railways, canals, and the Internet have not enjoyed in history. In both cases, supply first went out of control and then collapsed.

Every technological revolution in history has been caused by out-of-control supply. AI is the first to be forced to a rhythm by the laws of physics - this is its greatest luck.

A variable from space

Leave one long-term variable here: space data center.

The logic is science fiction but very hard - solar energy in sun-synchronous orbit is unlimited and free of charge; the satellite faces the deep sky at more than 200 degrees below zero, and the cost of heat dissipation is close to zero. The envisioned form is: the front end of the satellite is a solar panel, the middle is a standard server rack, and the tail is dragging a radiator hundreds of meters long. Multiple satellites are interconnected with lasers to form a virtual data center floating in orbit.

The two most expensive items in a ground data center—power and cooling—are free in space.

Timeline: Proof of concept may be seen within two years, and around 2030 it may begin to shake the investment logic of terrestrial data centers.

Remember this variable. It doesn't change anything yet, but it is a sword hanging over the entire L3 infrastructure layer - which will be used in a moment.

Where is the bubble really: demining layer by layer along the pyramid

After finishing talking about the two locks, go back to the five-layer map and go through each layer from bottom to top.

L0 + application layer header: Large cap - expensive, but not a bubble

Microsoft, Google, Meta, Amazon, Nvidia. Capital expenditures at this level correspond to real contracts, real revenue, and full load utilization.

Two numbers are enough.

The first one: AWS’s contracted but unfulfilled orders on hand reached US$360-370 billion in the first quarter, a year-on-year increase of more than 90%—this does not include the subsequent US$100 billion commitment of a leading AI laboratory. What concept? It's equivalent to AWS. Even if it doesn't sign a new customer from today on, the work it has signed will be enough for it to last for several years. These are not expectations, they are signed contracts.

Second: The leading large model company mentioned earlier - in 18 months, its revenue has increased from less than 100 million to more than 45 billion, and it has made a quarterly profit. There is no other example of this growth rate in the history of human business.

There is another account that few people calculate: the economics of reasoning. Training a cutting-edge model is a pure investment, and it costs money without blinking an eye; but after the model is trained, every time it is called and every time a token is generated, it is income. According to current industry estimates, the inference revenue opportunity throughout the life cycle of a model is approximately 5 to 10 times its pre-training investment. In other words, today's astronomical capital expenditures are not buying a one-time product like a "model", but a "computing power toll station" for many years to come.

The toll station model has a characteristic: the initial investment scares people to death, but the later cash flow drowns people. This is the case with highways, power grids, and telecommunications networks—provided there are actually cars running. And we've confirmed it: not a single GPU is dark, and every lane is full.

Is it expensive? expensive. Is it a bubble? The definition of a bubble is when prices diverge from fundamentals, and fundamentals are catching up with prices at a rate of 80x every 18 months.

Back then, the valuation stood still, waiting for income, until it went bankrupt; now, the income is chasing the valuation, and it can catch up.

To sum up the buyers at this level in one sentence: they are not betting on a story when buying computing power, but they have no choice in the face of the orders they have already received. Without expansion of production, the signed contracts cannot be delivered - this is capital expenditure driven by demand, not by illusion.

L1 memory layer: long and short strangulation area

Go up one level, memory chips. This is now the most acute battlefield for long and short differences.

First explain why this layer is important. If the GPU is the chef, the memory (especially the high-bandwidth memory HBM) is the food preparation table - no matter how fast the chef's knife skills are, it will be useless if the food cannot be delivered. AI inference is precisely a job that requires crazy "preparation speed": the larger the model, the longer the conversation, and the demand for memory bandwidth increases faster than the demand for computing power.

The current situation: Memory prices have increased by 60-70% in a year, and Micron's profit margin has soared from the historical average of 16% to 70%.

Put this number into history to see how scary it is: In the past twenty-five years, the memory industry has been famous for its "pig cycle" - rising prices, crazy production expansion, oversupply, price collapse, collective losses, and the cycle repeats. Every time a profit margin of 70% appears in this industry, it is followed by a funeral. According to the old script, it is time to clear the warehouse and run away.

But the logic of the bulls is: this time the demand is not to replenish inventory, but is structural. The demand for HBM in AI reasoning will continue to increase, and memory manufacturers have been taught by the cycle for 25 years. This expansion is extremely cautious - no one wants to be the one who crashes the price.

There is a structural change here that deserves to be mentioned separately: after twenty-five years of bloody reshuffle, there are only three players left in the global high-end memory. In the 1990s, there were more than 20 manufacturers in this industry, and no one could stop the price war. Today, the three oligarchs are looking at each other's production expansion plans across the Pacific Ocean, and no one wants to take the initiative. The oligopoly structure naturally comes with its own production capacity discipline - this is the strongest structural reason that "this production expansion will not get out of control" and is more reliable than any management statement.

And HBM is still quietly "squeezing" the production capacity of ordinary memory: the same production line, the output of wafers cut to HBM is much less than that of ordinary memory. The stronger the demand for HBM, the tighter the supply of ordinary memory, and the price of the entire industry is pushed up together - this is why even the price of ordinary memory modules in your computer is rising.

A more important number: Currently, only about 0.1% of the world’s population actually uses AI correctly. If this number moves towards 5% - that is, from "geek toys" to "daily tools for ordinary white-collar workers" - the memory demand ceiling is above the clouds.

The logic of the short sellers is equally hard: the current price increase is driven by the price itself, not by shipments - hoarding goods, being reluctant to sell, buying up but not buying down. This is a typical signal of supply and demand mismatch, not a sign of healthy demand.

A 70% profit margin is either the starting point of a new era or the culmination of an old script. The long bet is "this time is different" - and these five words happen to be the five most expensive words in the history of investment.

We do not draw any conclusions at this level. It's a gambling table, not a bubble, with real chips on both sides.

L2 interconnection layer: optical module - the smell of bubbles, start here

Finally we’ve reached the point where we really want to hit the blackboard. Also the only "hardware exception" on that map.

First take thirty seconds to explain what an optical module is. There are tens of thousands of GPUs in an AI data center. They do not do their own work. Instead, they exchange data at all times and collaborate to calculate the same model. The amount of "dialogue" between chips is so large that copper wires cannot handle it. Electrical signals must be converted into optical signals and transmitted through optical fibers. The small box responsible for "converting electricity to light and light to electricity" is the optical module.

GPU is the muscle, optical module is the blood vessel. The larger the cluster size, the higher the demand for interconnection between chips. Therefore, the more popular AI becomes, the crazier optical modules become. This industry logic is true. The entire optical module market is expected to grow by nearly 60% this year, and production capacity will indeed be "sold out by 2028."

The logic is true. But let's take a look at what the stock prices have done on a company-by-house basis.

The first company: Lumentum - the son of the last bubble and the leader of this bubble

This company makes lasers and optical components. To put it bluntly, it is the core "light source" in optical modules and optical communication systems. Its family history is worth pondering: its predecessor was one of the most famous star stocks in the optical communications bubble in 2000. The company's market value once reached hundreds of billions of dollars that year. After the bubble burst, it fell by 99%, becoming a standard illustration of the "infrastructure bubble" in textbooks. Lumentum was a spinoff from that company.

In the intervening twenty years, it lived a very ordinary life: it provided lasers for iPhone's face recognition and components for telecommunications networks. It was a typical "good but boring" hardware company.

Then the AI came. Data centers require a large number of high-speed lasers, and the new generation of "putting optical paths directly into switches" technology has pushed it to the center of the stage. Even NVIDIA has invested US$2 billion in real money. So:The share price has risen more than 10-fold in the past 12 months.

Is business getting better? It's really getting better. With orders lined up until 2028, this is the real deal. But put the two numbers together: Its revenue growth is expected to be in the tens of percent per year for the next few years, and the stock price is up more than a thousand percent in a year. The market’s pricing for it is already dozens of times its annual revenue—and the normal level for a mature hardware company is three to five times.

The center of the last burst of bubbles was light, and the strongest point of this bubble is still light. History doesn't repeat, but it does rhyme.

The second company: AAOI - a person who fell once stood on the same cliff again

This company makes complete optical transceiver modules, which are mainly sold to data centers of cloud manufacturers. Its history is also worth pondering: During the last wave of data center construction (around 2017), it was once a big bull stock - until its largest customer suddenly cut orders and switched to other suppliers. The stock price fell by 90% in the following two years, and then struggled on the edge of losses for seven or eight years.

Then AI came, demand for a new generation of high-speed optical modules exploded, and old customers came back. So:The share price increased more than 4 times during the year.

Pay attention to the difference between this company and Lumentum: Lumentum is an industry leader, has a technology moat, and is endorsed by NVIDIA; AAOI is a second-tier manufacturer that has not made money for most of the past decade, has a high concentration of customers, and has already learned a lesson from being cut off in the last round. Its surge is almost purely due to the buoyancy of the plate tide.

And the tide has begun to sway. Last month, there were more than one single-day double-digit plunge in this sector - AAOI fell more than 10% in one day, and the leading companies also fell 7%-10%. There is no real negative, but the high chips are starting to loosen.

There is another level of risk that is rarely discussed: the technical route itself.

Now the industry is promoting an architectural revolution: moving optical devices from "independent small boxes plugged into switches" directly into chip packages - called co-packaged optics in the industry. Once this direction becomes mainstream, it means two things: first, "optical modules" will be gradually absorbed as an independent product form, and the dominance will be transferred from module manufacturers to chip giants; second, the value in the chain will be concentrated in the "core light source", and the profits of the assembly process will be squeezed out.

Translation: For companies like Lumentum that hold lasers, this technological change brings more opportunities than risks - light sources will always be needed and are more valuable; but for module factories like AAOI that are good at assembly, it is a second knife hanging over their heads. Ironically, the market is now pricing the two types of companies with almost equal enthusiasm - when the tide is high, no one checks who is wearing swimming trunks.

In the same sector, some people sell irreplaceable light sources, while others sell boxes that may be bypassed by the architectural revolution at any time - but there is no difference in the stock price increase. This in itself is characteristic of a bubble.

To sum up the accounts at this level: demand has increased by nearly 60%, and the stock price has increased four to ten times. What's the gap in between? It is the market that discounts 2028 revenue into the 2026 stock price in advance.